Josh Barro observed on Twitter: “A single poll finding Trump up 30 in Oklahoma moves the national 538 forecast by 1.3 points. Seems… sensitive.”, and linked to 538’s page.
This got my attention, because I have been watching 538 a bit more closely than I’m proud of. I agree that this does seem sensitive. A lot of the comments here seem to explain it by saying that the move in Oklahoma implies moves in other more competitive states that Oklahoma results are historically correlated with (such as Texas?), and the model is picking up on this.
I had suspected this was just noise. FiveThirtyEight maintains three models: “polls-only” (which attempts to predict how the election will turn out on November 8 based only on polls), “polls-plus” (which incorporates demographics and fundamental), and “now-cast” (which attempts to predict how the election would turn out today).
Each of these results is based on 20,000 simulations of the election; thus we’re looking at differences between noisy data. However all three of fivethirtyeight’s models moved in the same direction at this time, according to the (unofficial) FiveThirtyEight forecast bot:
Polls-plus ↓ 1.3% (83.7%-16.2%)
Polls-only ↓ 1.4% (86.0%-13.9%)
Now-cast ↓ 1.9% (86.0%-13.9%)https://t.co/chUUwwNk1F
— 538 Forecast Bot (@538forecastbot) October 23, 2016
So what’s going on? To take a quick look at this, let’s look at the last 50 updates in which all three forecasts moved (sometimes there’s a change of 0.0). By my count, 29 of them had all three forecasts moving in the same direction, while the forecasts have been roughly flat over that time. If the forecasts were uncorrelated, then we could think of the direction of the moves as just random flips of a coin, and one-fourth of the moves would have either three ups or three downs. So by this simple test, the moves appear to be correlated. In the spirit of Averaging All The Things that pervades FiveThirtyEight, perhaps the “true” motion was the average of the motion in the three models, 1.5 or 1.6 percent down.
So it seems quite reasonable that that move is “real” – at the very least all three models bear show it. But why? There’s at least one large move based on other single polls. Looking at the history of updates, we find this move, less than an hour ago as I write this, based on a poll that shows Nevada at Clinton +6 when the model had Clinton up by 4 previously:
Polls-plus ↑ 1.0% (84.0%-16.0%)
Polls-only ↑ 1.0% (86.1%-13.9%)
Now-cast ↑ 1.3% (86.6%-13.3%)https://t.co/chUUwwNk1F
— 538 Forecast Bot (@538forecastbot) October 24, 2016
But most big moves come from several polls coming in at the same time, and in more significant states. I think the real lesson here is that obsessive poll-watching is exhausting…